Likelihood Training of Schrödinger Bridge Using Forward-Backward SDEs Theory

Explore the novel computational framework for likelihood training of Schrödinger Bridge models in this comprehensive lecture. Delve into the theory of Forward-Backward Stochastic Differential Equations and its application to generative modeling. Learn how this approach generalizes Score-based Generative Models and enables the use of modern generative training techniques. Discover the optimization principles behind Schrödinger Bridge models and their potential for generating realistic images. Follow along as the speaker covers topics such as Deep Generalized Schrödinger Bridge, Schrödinger Bridge Theory, Log-Likelihood as Path Integral, Mean-Field Games, and practical solutions for Deep Generalized Schrödinger Bridge. Gain insights into the comparative results on image generation for datasets like MNIST, CelebA, and CIFAR10.

- Introduction
- Deep Generalized Schrödinger Bridge
- Schrödinger Bridge Theory
- Log-Likelihood as Path Integral
- Mean-Field Games
- Solving Deep Generalized Schrödinger Bridge & Results
- Summary
- Q+A